NVIDIA Nemotron-3.5 Bolt Tested: Strong Vision, Shaky Coding
Hands-on tests of NVIDIA Nemotron-3.5 Bolt on bug fixing, creative coding, vision reasoning, and translation show a sharp but uneven model.
What is NVIDIA Nemotron-3.5 Bolt?
Nemotron-3.5 Bolt is a vision-language model from NVIDIA built with 74 billion total parameters, of which only 9 billion are active per token through a mixture-of-experts design. It’s a hybrid of Mamba-2 and Transformer architectures, compressed down from a larger 120-billion-parameter sibling and then quantized to 2-bit precision using quantization-aware training (QAT). The resulting model file lands around 22 GB, small enough to run on a single GPU with 24 to 48 GB of VRAM while leaving room for other applications. It handles text, images, and video, with a context window NVIDIA lists at up to 256K tokens. In independent hands-on testing, the model showed real strength in vision reasoning and scientific diagrams, but struggled with agentic coding and multilingual translation.
TL;DR
- Nemotron-3.5 Bolt is a 74B-parameter MoE vision-language model with only 9B active parameters per token, quantized to 2-bit and packaged at roughly 22 GB, designed to run locally on 24 to 48 GB of VRAM.
- In a local bug-fixing test on a full-stack stud farm management app (Postgres, Flask, Nginx, Redis via Docker Compose), the model failed to locate a hardcoded gestation-length error and instead validated the broken logic as correct.
- A creative coding task asking for an animated doner kebab skewer rotating in front of a gas heater, built as a single HTML file with no libraries, produced clean-looking reasoning but a broken visual output, even after the model was shown a screenshot of its own mistake.
- The model delivered a near-perfect vision reasoning result, correctly tracking a laptop’s location through a multi-character story with conflicting beliefs about where the object actually was, scoring nine out of nine correct answers.
- A scientific reasoning test involving a hand-drawn circuit diagram with intentional errors (a mislabeled resistor, a dangling wire, a meter with no leads) was solved correctly across all seven calculated values.
- Translation performance was inconsistent: major world languages came through reasonably, but the model hallucinated or looped on less common ones when translating a Turkish news headline into more than 80 languages.
- Taken together, the results suggest Nemotron-3.5 Bolt is a capable reasoning and vision model but not yet reliable enough for production coding tasks without human review.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How was Nemotron-3.5 Bolt tested?
The testing used a local deployment via llama.cpp with speculative decoding powered by a small drafter model, running on a single GPU consuming close to 30 GB of VRAM. This setup mirrors how a developer with a high-end consumer or prosumer GPU might realistically run the model at home or on a rented cloud GPU instance, rather than relying on a hosted API.
Four distinct task types were used to probe different capabilities:
- Debugging a real application. A working stud farm management app with a hidden logic bug (a hardcoded gestation period of 330 days instead of the correct ~340) was handed to the model with no hints about where the problem lived.
- Creative coding from a text brief. The model was asked to generate a single self-contained HTML file animating a rotating kebab skewer in front of a heater, using plain canvas code with no external libraries.
- Vision reasoning over a narrative. An AI-generated image of a coffee shop scene was paired with a multi-step story about a laptop moving between people, each of whom believed it was somewhere different. The model had to reconstruct the real sequence of events from the image and prompt.
- Scientific diagram reading. A hand-drawn circuit with an intentionally wrong label, a dangling wire, and a voltmeter with no leads attached was given to the model to calculate real circuit values rather than trust the (incorrect) on-image annotations.
A fifth task tested multilingual translation of a news headline from a Turkish newspaper into over 80 languages in a single pass.
Where did Nemotron-3.5 Bolt fail?
The clearest failure came in the bug-fixing test. The app in question was not a toy; it ran on a real stack with a Postgres database, a Flask API backend, an Nginx frontend, and Redis, all orchestrated through Docker Compose. The bug itself was subtle but consequential: a hardcoded assumption of 330 days of horse gestation instead of the real-world average of roughly 340, which pushed every due date forward by ten days and incorrectly flagged a mare as overdue in the dashboard.
Despite being given the goal of finding and fixing the issue without hints, the model’s reasoning looped without ever questioning the 330-day assumption. It “verified” the flawed logic as correct, reorganized unrelated parts of the code (grouping, database queries), and left the actual bug untouched. A hard refresh of the running app confirmed the dashboard still showed the wrong status.
The creative coding test produced a similar pattern: thoughtful-looking reasoning about draw order and overlap checks, clean-looking code, but a broken final render. Giving the model a screenshot of its own failed output and asking it to compare that against the original brief (testing its multimodal self-correction) didn’t fix the problem either. The second generation attempt still didn’t match what was asked for.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Translation was the other weak spot. While mainstream languages translated reasonably well, the model began hallucinating content and effectively looping once it moved into less common languages, undermining confidence in the 80-plus language batch output.
Where did Nemotron-3.5 Bolt succeed?
The vision reasoning and scientific reasoning tests told a very different story. In the laptop-tracking scenario, the model had to infer, from a single image and a layered narrative, who actually moved a laptop and why three different characters ended up believing it was in three different places. It answered all nine comprehension questions correctly, including tricky ones involving overlapping beliefs between characters and a visually similar object (a silver laptop) causing confusion in the story.
The circuit diagram test was arguably the strongest result. The hand-drawn image deliberately included a wrong label, a wire that led nowhere, and a voltmeter reading with no leads drawn, designed to catch any model that pattern-matches on labels instead of reasoning about actual circuit topology. Nemotron-3.5 Bolt correctly identified that the labeled resistor values were wrong, treated the dangling wire as carrying no current, and calculated all seven required values correctly.
Is Nemotron-3.5 Bolt worth using?
For developers specifically interested in visual reasoning, document and diagram understanding, or multimodal step-by-step problem solving, Nemotron-3.5 Bolt performed at a level the testing described as “world-class” on some answers, which is notable for a heavily quantized 2-bit model designed to run on a single consumer-grade GPU.
For coding, especially agentic or open-ended tasks like locating a non-obvious bug in a multi-service application or generating correct creative frontend code from a brief, the results were weaker. The model’s compressed, highly efficient design (9B active parameters out of 74B total, 2-bit quantization) trades off some precision, and that tradeoff appears to show up specifically in tasks that require sustained logical tracking across a codebase rather than a single image or diagram.
The practical takeaway for builders: Nemotron-3.5 Bolt looks like a strong candidate for local, GPU-constrained vision and reasoning workloads, but it is not yet a drop-in replacement for a human reviewing production code changes.
Frequently Asked Questions
What hardware do you need to run Nemotron-3.5 Bolt locally?
The quantized model file is around 22 GB and NVIDIA designed it to fit in 24 to 48 GB of VRAM, leaving headroom to run alongside other applications on a single GPU.
Is Nemotron-3.5 Bolt good at coding?
In hands-on testing it struggled with both a real-world bug-fixing task and a creative HTML/canvas coding task, failing to identify a hidden logic bug and producing a broken animated output even after a correction attempt. It should not be trusted unsupervised for production coding work.
How does Nemotron-3.5 Bolt perform on vision tasks?
It performed strongly, correctly answering all nine questions in a complex multi-character vision reasoning test and solving a hand-drawn circuit diagram problem with deliberately misleading labels, getting every calculated value right.
Can Nemotron-3.5 Bolt translate text reliably?
Performance was mixed. Major languages translated acceptably, but when pushed to translate a headline into more than 80 languages, the model hallucinated content and appeared to loop on less common languages.
What makes Nemotron-3.5 Bolt different architecturally?
It combines a Mamba-2 and Transformer hybrid design with a mixture-of-experts setup (74B total parameters, 9B active per token), compressed from a larger 120B model and quantized to 2-bit using quantization-aware training, plus a small drafter model for speculative decoding.
